Bunch 转换为 HDF5 文件：高效存储 Cifar 等数据集

关于如何将数据集封装为 Bunch 可参考 关于 『AI 专属数据库的定制』的改进。

PyTables 是 Python 与 HDF5 数据库/文件标准的结合。它专门为优化 I/O 操作的性能、最大限度地利用可用硬件而设计，并且它还支持压缩功能。

下面的代码均是在 Jupyter NoteBook 下完成的：

import sys
sys.path.append('E:/xinlib')
from base.filez import DataBunch

import tables as tb
import numpy as np

def bunch2hdf5(root):
    '''
    这里我仅仅封装了 Cifar10、Cifar100、MNIST、Fashion MNIST 数据集，
    使用者还可以自己追加数据集。
    '''
    db = DataBunch(root)
    filters = tb.Filters(complevel=7, shuffle=False)
    # 这里我采用了压缩表，因而保存为 `.h5c` 但也可以保存为 `.h5`
    with tb.open_file(f'{root}X.h5c', 'w', filters=filters, title='Xinet\'s dataset') as h5:
        for name in db.keys():
            h5.create_group('/', name, title=f'{db[name].url}')
            if name != 'cifar100':
                h5.create_array(h5.root[name], 'trainX', db[name].trainX, title='训练数据')
                h5.create_array(h5.root[name], 'trainY', db[name].trainY, title='训练标签')
                h5.create_array(h5.root[name], 'testX', db[name].testX, title='测试数据')
                h5.create_array(h5.root[name], 'testY', db[name].testY, title='测试标签')
            else:
                h5.create_array(h5.root[name], 'trainX', db[name].trainX, title='训练数据')
                h5.create_array(h5.root[name], 'testX', db[name].testX, title='测试数据')
                h5.create_array(h5.root[name], 'train_coarse_labels', db[name].train_coarse_labels, title='超类训练标签')
                h5.create_array(h5.root[name], 'test_coarse_labels', db[name].test_coarse_labels, title='超类测试标签')
                h5.create_array(h5.root[name], 'train_fine_labels', db[name].train_fine_labels, title='子类训练标签')
                h5.create_array(h5.root[name], 'test_fine_labels', db[name].test_fine_labels, title='子类测试标签')

        for k in ['cifar10', 'cifar100']:
            for name in db[k].meta.keys():
                name = name.decode()
                if name.endswith('names'):
                    label_names = np.asanyarray([label_name.decode() for label_name in db[k].meta[name.encode()]])
                    h5.create_array(h5.root[k], name, label_names, title='标签名称')

完成 Bunch到 HDF5 的转换

root = 'E:/Data/Zip/'
bunch2hdf5(root)

h5c = tb.open_file('E:/Data/Zip/X.h5c')
h5c

File(filename=E:/Data/Zip/X.h5c, title="Xinet's dataset", mode='r', root_uep='/', filters=Filters(complevel=7, complib='zlib', shuffle=False, bitshuffle=False, fletcher32=False, least_significant_digit=None))
/ (RootGroup) "Xinet's dataset"
/cifar10 (Group) 'https://www.cs.toronto.edu/~kriz/cifar.html'
/cifar10/label_names (Array(10,)) '标签名称'
  atom := StringAtom(itemsize=10, shape=(), dflt=b'')
  maindim := 0
  flavor := 'numpy'
  byteorder := 'irrelevant'
  chunkshape := None
/cifar10/testX (Array(10000, 32, 32, 3)) '测试数据'
  atom := UInt8Atom(shape=(), dflt=0)
  maindim := 0
  flavor := 'numpy'
  byteorder := 'irrelevant'
  chunkshape := None
/cifar10/testY (Array(10000,)) '测试标签'
  atom := Int32Atom(shape=(), dflt=0)
  maindim := 0
  flavor := 'numpy'
  byteorder := 'little'
  chunkshape := None
/cifar10/trainX (Array(50000, 32, 32, 3)) '训练数据'
  atom := UInt8Atom(shape=(), dflt=0)
  maindim := 0
  flavor := 'numpy'
  byteorder := 'irrelevant'
  chunkshape := None
/cifar10/trainY (Array(50000,)) '训练标签'
  atom := Int32Atom(shape=(), dflt=0)
  maindim := 0
  flavor := 'numpy'
  byteorder := 'little'
  chunkshape := None
/cifar100 (Group) 'https://www.cs.toronto.edu/~kriz/cifar.html'
/cifar100/coarse_label_names (Array(20,)) '标签名称'
  atom := StringAtom(itemsize=30, shape=(), dflt=b'')
  maindim := 0
  flavor := 'numpy'
  byteorder := 'irrelevant'
  chunkshape := None
/cifar100/fine_label_names (Array(100,)) '标签名称'
  atom := StringAtom(itemsize=13, shape=(), dflt=b'')
  maindim := 0
  flavor := 'numpy'
  byteorder := 'irrelevant'
  chunkshape := None
/cifar100/testX (Array(10000, 32, 32, 3)) '测试数据'
  atom := UInt8Atom(shape=(), dflt=0)
  maindim := 0
  flavor := 'numpy'
  byteorder := 'irrelevant'
  chunkshape := None
/cifar100/test_coarse_labels (Array(10000,)) '超类测试标签'
  atom := Int32Atom(shape=(), dflt=0)
  maindim := 0
  flavor := 'numpy'
  byteorder := 'little'
  chunkshape := None
/cifar100/test_fine_labels (Array(10000,)) '子类测试标签'
  atom := Int32Atom(shape=(), dflt=0)
  maindim := 0
  flavor := 'numpy'
  byteorder := 'little'
  chunkshape := None
/cifar100/trainX (Array(50000, 32, 32, 3)) '训练数据'
  atom := UInt8Atom(shape=(), dflt=0)
  maindim := 0
  flavor := 'numpy'
  byteorder := 'irrelevant'
  chunkshape := None
/cifar100/train_coarse_labels (Array(50000,)) '超类训练标签'
  atom := Int32Atom(shape=(), dflt=0)
  maindim := 0
  flavor := 'numpy'
  byteorder := 'little'
  chunkshape := None
/cifar100/train_fine_labels (Array(50000,)) '子类训练标签'
  atom := Int32Atom(shape=(), dflt=0)
  maindim := 0
  flavor := 'numpy'
  byteorder := 'little'
  chunkshape := None
/fashion_mnist (Group) 'https://github.com/zalandoresearch/fashion-mnist'
/fashion_mnist/testX (Array(10000, 28, 28, 1)) '测试数据'
  atom := UInt8Atom(shape=(), dflt=0)
  maindim := 0
  flavor := 'numpy'
  byteorder := 'irrelevant'
  chunkshape := None
/fashion_mnist/testY (Array(10000,)) '测试标签'
  atom := Int32Atom(shape=(), dflt=0)
  maindim := 0
  flavor := 'numpy'
  byteorder := 'little'
  chunkshape := None
/fashion_mnist/trainX (Array(60000, 28, 28, 1)) '训练数据'
  atom := UInt8Atom(shape=(), dflt=0)
  maindim := 0
  flavor := 'numpy'
  byteorder := 'irrelevant'
  chunkshape := None
/fashion_mnist/trainY (Array(60000,)) '训练标签'
  atom := Int32Atom(shape=(), dflt=0)
  maindim := 0
  flavor := 'numpy'
  byteorder := 'little'
  chunkshape := None
/mnist (Group) 'http://yann.lecun.com/exdb/mnist'
/mnist/testX (Array(10000, 28, 28, 1)) '测试数据'
  atom := UInt8Atom(shape=(), dflt=0)
  maindim := 0
  flavor := 'numpy'
  byteorder := 'irrelevant'
  chunkshape := None
/mnist/testY (Array(10000,)) '测试标签'
  atom := Int32Atom(shape=(), dflt=0)
  maindim := 0
  flavor := 'numpy'
  byteorder := 'little'
  chunkshape := None
/mnist/trainX (Array(60000, 28, 28, 1)) '训练数据'
  atom := UInt8Atom(shape=(), dflt=0)
  maindim := 0
  flavor := 'numpy'
  byteorder := 'irrelevant'
  chunkshape := None
/mnist/trainY (Array(60000,)) '训练标签'
  atom := Int32Atom(shape=(), dflt=0)
  maindim := 0
  flavor := 'numpy'
  byteorder := 'little'
  chunkshape := None

从上面的结构可看出我将 Cifar10、Cifar100、MNIST、Fashion MNIST 进行了封装，并且还附带了它们各种的数据集信息。比如标签名，数字特征（以数组的形式进行封装）等。

%%time
arr = h5c.root.cifar100.trainX.read() # 读取数据十分快速

Wall time: 125 ms

arr.shape

(50000, 32, 32, 3)

h5c.root

/ (RootGroup) "Xinet's dataset"
  children := ['cifar10' (Group), 'cifar100' (Group), 'fashion_mnist' (Group), 'mnist' (Group)]

X.h5c 使用说明

下面我们以 Cifar100 为例来展示我们自创的数据集 X.h5c（我将其上传到了百度云盘「链接：https://pan.baidu.com/s/1hsbMhv3MDlOES3UDDmOQiw 密码：qlb7」可以下载直接使用；亦可你自己生成，不过我推荐自己生成，可以对数据集加深理解）

cifar100 = h5c.root.cifar100
cifar100

/cifar100 (Group) 'https://www.cs.toronto.edu/~kriz/cifar.html'
  children := ['coarse_label_names' (Array), 'fine_label_names' (Array), 'testX' (Array), 'test_coarse_labels' (Array), 'test_fine_labels' (Array), 'trainX' (Array), 'train_coarse_labels' (Array), 'train_fine_labels' (Array)]

'coarse_label_names' 指的是粗粒度或超类标签名，'fine_label_names' 则是细粒度标签名。

可以使用 read() 方法直接获取信息，也可以使用索引的方式获取。

coarse_label_names = cifar100.coarse_label_names[:]
# 或者
coarse_label_names = cifar100.coarse_label_names.read()
coarse_label_names.astype('str')

array(['aquatic_mammals', 'fish', 'flowers', 'food_containers',
       'fruit_and_vegetables', 'household_electrical_devices',
       'household_furniture', 'insects', 'large_carnivores',
       'large_man-made_outdoor_things', 'large_natural_outdoor_scenes',
       'large_omnivores_and_herbivores', 'medium_mammals',
       'non-insect_invertebrates', 'people', 'reptiles', 'small_mammals',
       'trees', 'vehicles_1', 'vehicles_2'], dtype='<U30')

fine_label_names = cifar100.fine_label_names[:].astype('str')
fine_label_names

array(['apple', 'aquarium_fish', 'baby', 'bear', 'beaver', 'bed', 'bee',
       'beetle', 'bicycle', 'bottle', 'bowl', 'boy', 'bridge', 'bus',
       'butterfly', 'camel', 'can', 'castle', 'caterpillar', 'cattle',
       'chair', 'chimpanzee', 'clock', 'cloud', 'cockroach', 'couch',
       'crab', 'crocodile', 'cup', 'dinosaur', 'dolphin', 'elephant',
       'flatfish', 'forest', 'fox', 'girl', 'hamster', 'house',
       'kangaroo', 'keyboard', 'lamp', 'lawn_mower', 'leopard', 'lion',
       'lizard', 'lobster', 'man', 'maple_tree', 'motorcycle', 'mountain',
       'mouse', 'mushroom', 'oak_tree', 'orange', 'orchid', 'otter',
       'palm_tree', 'pear', 'pickup_truck', 'pine_tree', 'plain', 'plate',
       'poppy', 'porcupine', 'possum', 'rabbit', 'raccoon', 'ray', 'road',
       'rocket', 'rose', 'sea', 'seal', 'shark', 'shrew', 'skunk',
       'skyscraper', 'snail', 'snake', 'spider', 'squirrel', 'streetcar',
       'sunflower', 'sweet_pepper', 'table', 'tank', 'telephone',
       'television', 'tiger', 'tractor', 'train', 'trout', 'tulip',
       'turtle', 'wardrobe', 'whale', 'willow_tree', 'wolf', 'woman',
       'worm'], dtype='<U13')

'testX' 与 'trainX' 分别代表数据的测试数据和训练数据，而其他的节点所代表的含义也是类似的。

例如，我们可以看看训练集的数据和标签：

trainX = cifar100.trainX
train_coarse_labels = cifar100.train_coarse_labels

array([11, 15,  4, ...,  8,  7,  1])

shape 为 (50000, 32, 32, 3)，数据的获取，我们一样可以采用索引的形式或者使用 read()：

train_data = trainX[:]
print(train_data[0].shape)
print(train_data.dtype)

(32, 32, 3)
uint8

当然，我们也可以直接使用 trainX 做运算。

for x in cifar100.trainX:
    y = x * 2
    break

print(y.shape)

(32, 32, 3)

h5c.get_node(h5c.root.cifar100, 'trainX')

/cifar100/trainX (Array(50000, 32, 32, 3)) '训练数据'
  atom := UInt8Atom(shape=(), dflt=0)
  maindim := 0
  flavor := 'numpy'
  byteorder := 'irrelevant'
  chunkshape := None

更甚者，我们可以直接定义迭代器来获取数据：

trainX = cifar100.trainX
train_coarse_labels = cifar100.train_coarse_labels

def data_iter(X, Y, batch_size):
    n = X.nrows
    idx = np.arange(n)
    if X.name.startswith('train'):
        np.random.shuffle(idx)
    for i in range(0, n ,batch_size):
        k = idx[i: min(n, i + batch_size)].tolist()
        yield np.take(X, k, 0), np.take(Y, k, 0)

for x, y in data_iter(trainX, train_coarse_labels, 8):
    print(x.shape, y)
    break

(8, 32, 32, 3) [ 7  7  0 15  4  8  8  3]

更多使用详情见：使用 迭代器 获取 Cifar 等常用数据集